Do We Need Anisotropic Graph Neural Networks?

Overview

This repo provides the implementation of the Efficient Graph Convolution (EGC) layer using PyTorch Geometric. We include hyperparameters and pretrained models (even for baselines which we trained!).

• hyperparameters.md contains hyperparameters and experiment details.
• experiments/layers.py contains the layer definitions for EGC, and is likely what you are most interested in. There's also an optimized implementation that was upstreamed to PyTorch Geometric in experiments/optimized_layers.py.
• The experiments directory contains subdirectories for each experiment dataset.
• The code is structured using my experiment tuning library (exptune) which enables fast hyperparameter search + training of final models + plotting and so on. This is bundled under the third_party directory as a git submodule. Beyond these dependencies, we use PyTorch Geometric and OGB.
• To retrain the main table + ablation studies, you can use train_main_table.sh and train_ablation.sh.
• To run the pre-trained models, use run_pretrained.sh.

Getting Started

Clone the repo with the submodules (--recurse-submodules) and build the Dockerfile:

# Clone
git clone https://github.com/shyam196/egc.git --recurse-submodules
# Build
docker build . -t camlsys/egc
# Run
docker run -it --gpus all camlsys/egc bash

Using Pre-Trained Models

The script run_pretrained.sh provides all the commands you need to download the models and perform a run over the test set.

Re-Generating Experiment Results

We provide two scripts to generate the results for the main table, and for the ablation studies. This is discussed in more depth in hyperparameters.md.

Citing This Work

@inproceedings{
tailor2022egc,
title={Do We Need Anistropic Graph Neural Networks?},
author={Shyam A. Tailor and Felix Opolka and Pietro Lio and Nicholas Donald Lane},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=hl9ePdHO4_s}
}